AI Expense Management

Because it is not a few emails — it is a few emails per unsubstantiated transaction, per client, every month, forever, and each one is socially awkward enough that it does not get sent. That is the whole mechanism. Nobody minds sending the first reminder. The fourth reminder, to a client who is paying you and has ignored the first three, is unpleasant, so it gets deferred, and the transaction quietly enters the BAS quarter without support. By the time anyone looks, the client cannot remember what the $340 at a hardware store in March was for either. The value of automating this is not the minutes saved. It is that the fourth reminder actually gets sent, unemotionally, with the date, amount and merchant attached — which is usually what makes the client able to answer at all.

No, and this is one of the clearest cases where it should not try. Whether a purchase was for the business, what proportion of a phone bill is private, whether a meal was entertainment, whether a vehicle expense needs apportioning — these turn on facts that exist only in the client’s head and in the substantiating evidence, and on professional judgement about the tax treatment. A model looking at a card feed line reading "COLES 4471" has no way to know if it was staff amenities or the family shop. What it does instead is useful and honest: it recognises which transactions historically involve an apportionment or a private-use question in that file, flags them with that history attached, and asks the client the specific question — then puts the answer in front of your team to determine the treatment.

Flagged, never decided. Entertainment and fringe benefits are among the most judgement-heavy areas in a small business file: the treatment depends on who was present, what the purpose was, where it happened, and which method the client uses — none of which is visible in a receipt total. What the AI can do is recognise the categories of spending that commonly raise these questions — restaurants, functions, travel, client gifts, motor vehicles, staff benefits — and surface them for your team with the transaction history and the substantiating document attached, rather than coding them to Entertainment and moving on. The determination is a professional judgement for your registered agent or accountant, and the deliberate absence of an AI opinion at that point is a feature.

Card feed lines and receipts are matched against each other, and each side chases the other. When a card transaction appears with no receipt, the AI identifies the cardholder and asks them for it directly, with the merchant, date and amount — repeatedly, on your schedule, without needing anyone at the firm to feel awkward about it. When a receipt is submitted with no corresponding card line, it is held until the transaction appears rather than double-counted. Where a receipt covers several card lines or one card line covers several receipts, it reconciles the group. What matters is that the gap between the two is always visible and always shrinking, instead of being discovered as a list of forty unsubstantiated transactions the week before the BAS is due.

It matters, and we would rather be straight about it. A crumpled thermal receipt photographed in a dim car park is genuinely hard to read, and thermal paper fades to blank within months, which is a records problem no software solves. The AI reads what is readable — merchant, date, total, GST, and line items where they are legible — and where a document is too degraded to read with confidence, it says so and asks the client rather than producing plausible numbers. That behaviour is the important part: a wrong read that announces itself costs thirty seconds, and a wrong read that does not can flow into a BAS. The genuinely useful habit shift is capturing the receipt at the point of sale rather than in a shoebox in July, and continuous chasing is what actually produces that habit.

Digital copies are generally acceptable to the ATO provided they are a true and clear reproduction of the original, and records generally need to be kept for five years — but the specifics of what your clients must retain, and for how long in their particular circumstances, is a question for your registered agent rather than for us or for the software. What we can say about the product is what it does: the original image is retained and attached to the transaction it supports, alongside the extracted data and the review trail. The common failure in firms is not losing documents but severing them from the transactions they substantiate, so that nobody can produce the right one when it is actually needed. Keeping them together is the point.